An algorithmic co-scheduling method and system fusing risk quantification and multiple incentives

Abstract

The application provides an algorithm and electricity collaborative scheduling method and system fusing risk quantification and multiple incentives, the method comprises the following steps: acquiring multi-source data of a region where a data center is located; the multi-source data comprises wind and light prediction data, carbon price signals, electricity price signals, peak regulation compensation prices, standby capacity prices, algorithm power task queues and energy storage states; based on the multi-source data, the conditional value at risk method is used to calculate expected economic losses possibly caused by wind and light prediction errors in future time periods; according to recent prediction error characteristics and carbon price signals and peak regulation compensation prices, a risk preference coefficient is adaptively adjusted; based on the expected economic losses possibly caused by wind and light prediction errors in future time periods and the adjusted risk preference coefficient, a comprehensive cost objective function is constructed; the comprehensive cost objective function is solved under preset constraint conditions with the aim of minimizing total cost, so that an optimal scheduling scheme is obtained and output, and the reliability of algorithm and electricity collaborative scheduling is improved.

Description

Technical Field

[0001] This invention relates to the field of collaborative scheduling of power systems and computing resources, and in particular to a collaborative scheduling method and system that integrates risk quantification and multiple incentives. Background Technology

[0002] The strong intermittency and volatility of wind and solar power pose a severe challenge to the power supply reliability of data centers. When wind and solar power output drops sharply, if energy storage is insufficient or dispatch response is delayed, it can lead to minor issues such as interruption of computing tasks and breach of service level agreements (SLAs), or even serious issues such as equipment downtime and data loss.

[0003] Currently, data centers mainly face the following technical bottlenecks in the coordinated dispatch of wind, solar, energy storage, and power grid: (1) The lack of precise risk quantification methods for wind and solar forecasting errors. Existing dispatching systems generally use point forecasting methods to deal with the uncertainties of wind and solar power, but these methods can only provide a "possible range" and cannot quantify how much economic loss will be caused in extreme cases. For example, when the weather forecast shows "there may be clouds in the afternoon", traditional methods will only raise the lower limit of energy storage reserve capacity (such as from 10% to 20%), but cannot answer "how much additional electricity purchase cost will be required if the photovoltaic power drops from 100kW to 30kW". This deficiency makes it difficult to balance dispatching decisions between "too risky" (possible power outage) and "too conservative" (missing out on green electricity benefits).

[0004] (2) Carbon emission costs and peak-shaving market revenues are not included in unified optimization. On the one hand, the national carbon market is now officially operational, and the carbon emission costs (carbon price × carbon emission) of data centers purchasing electricity from the grid are becoming an important part of their operating costs. However, most existing dispatching methods only focus on electricity costs or green electricity consumption rates, without considering real-time carbon prices as optimization variables. On the other hand, data centers have a natural ability to adjust loads (delayed computing tasks, energy storage systems), and can obtain compensation by participating in the grid peak-shaving ancillary services market. However, existing dispatching systems generally treat data centers as pure electricity users and do not have a mechanism to actively utilize peak-shaving revenues. The lack of these two economic factors (carbon costs + peak-shaving revenues) means that data centers may lose hundreds of thousands to millions of yuan in potential benefits annually.

[0005] (3) Fixed risk appetite, unable to adapt to weather and market dynamics. The reliability of wind and solar forecasts varies greatly under different weather conditions: forecast errors are small under clear weather, allowing for aggressive strategies (leaving less energy storage for backup, using more wind and solar power, or even applying for more peak-shaving capacity for arbitrage); while forecast errors are large under cloudy / rainy / windy weather, requiring a conservative strategy (reserving more emergency energy storage capacity). However, the risk appetite of existing dispatch systems (such as the lower limit of energy storage SOC) is often a pre-set fixed value, unable to be dynamically adjusted according to real-time forecast error characteristics (average error, error change rate) and market price signals (carbon price, peak-shaving compensation). This leads to excessive conservatism and loss of benefits under clear weather, while being unable to switch to conservatism in time when the weather changes suddenly, thus facing the risk of power outages.

[0006] (4) Lack of carbon emission traceability for energy storage and crude carbon cost calculation. Energy storage systems may absorb both wind and solar green electricity and grid thermal power during charging, but lack accurate traceability of carbon sources during discharging. The traditional approach simply calculates the carbon emissions from energy storage discharge based on the grid average carbon emission factor. This overestimates the carbon reduction contribution of green electricity (when the proportion of green electricity in energy storage charging is high) and underestimates carbon costs (when the proportion of thermal power in energy storage charging is high). Inaccurate carbon cost measurement can cause dispatch decisions to deviate from the optimal low-carbon path.

[0007] In summary, existing technologies have significant shortcomings in handling wind and solar uncertainties, carbon markets, peak-shaving markets, and risk self-adaptation. There is an urgent need for a computing-coordinated scheduling method that can integrate the quantification of wind and solar uncertainty risks, carbon emission costs, peak-shaving ancillary service revenues, and dynamically and adaptively adjust risk preferences. Summary of the Invention

[0008] To address the problems existing in the prior art, this invention innovatively proposes a computing-powered collaborative scheduling method and system that integrates risk quantification and multiple incentives. This effectively solves the problem of low reliability in computing-powered collaborative scheduling caused by the prior art and significantly improves the reliability of computing-powered collaborative scheduling.

[0009] The first aspect of this invention provides a computing-electricity collaborative scheduling method that integrates risk quantification and multiple incentives, comprising: Acquire multi-source data of the region where the data center is located; the multi-source data includes wind and solar forecast data, carbon price signals, electricity price signals, peak-shaving compensation prices, standby capacity prices, computing power task queues, and energy storage status; Based on multi-source data, the conditional value at risk method is used to calculate the expected economic losses that may be caused by wind and solar forecasting errors in different future periods. Based on recent forecast error characteristics, carbon price signals, and peak-shaving compensation prices, the risk preference coefficient is adaptively adjusted. Based on the expected economic losses that may be caused by wind and solar forecast errors in different future periods and the adjusted risk appetite coefficient, a comprehensive cost objective function is constructed. With the goal of minimizing total cost, the comprehensive cost objective function is solved under preset constraints to obtain and output the optimal scheduling scheme.

[0010] A second aspect of the present invention provides a computing-electricity collaborative scheduling system that integrates risk quantification and multiple incentives, comprising: The acquisition module acquires multi-source data from the region where the data center is located; the multi-source data includes wind and solar forecast data, carbon price signals, electricity price signals, peak-shaving compensation prices, standby capacity prices, computing power task queues, and energy storage status; The calculation module, based on multi-source data, uses the conditional value at risk method to calculate the expected economic losses that may be caused by wind and solar forecast errors in different future periods. The adjustment module adaptively adjusts the risk preference coefficient based on recent forecast error characteristics, carbon price signals, and peak-shaving compensation prices. The module constructs a comprehensive cost objective function based on the expected economic losses that may be caused by wind and solar forecast errors in future time periods and the adjusted risk preference coefficient. The solution module aims to minimize the total cost by solving the comprehensive cost objective function under preset constraints, and then obtains and outputs the optimal scheduling scheme.

[0011] The technical solution adopted in this invention has the following technical effects: 1. The technical solution of this invention acquires multi-source data of the region where the data center is located; the multi-source data includes wind and solar forecast data, carbon price signals, electricity price signals, peak shaving compensation prices, reserve capacity prices, computing power task queues, and energy storage status; based on the multi-source data, the conditional value at risk method is used to calculate the expected economic losses that may be caused by wind and solar forecast errors in each future period; based on the recent forecast error characteristics and carbon price signals and peak shaving compensation prices, the risk preference coefficient is adaptively adjusted; based on the expected economic losses that may be caused by wind and solar forecast errors in each future period and the adjusted risk preference coefficient, a comprehensive cost objective function is constructed; with the goal of minimizing the total cost, the comprehensive cost objective function is solved under preset constraints to obtain and output the optimal scheduling scheme, effectively solving the problem of low reliability of computing-power collaborative scheduling caused by existing technologies, and effectively improving the reliability of computing-power collaborative scheduling.

[0012] 2. The technical solution of this invention introduces a risk preference coefficient, which automatically adjusts the conservative / aggressive level of risk preference based on the statistical characteristics and distribution characteristics of recent prediction errors and market price signals, thereby achieving a balance between risk and return.

[0013] 3. The risk loss function in the technical solution of this invention is not limited to a single power shortage penalty cost, but is constructed by comprehensively considering different dimensions such as direct economic losses, computing power service default losses, and accelerated energy storage life loss, thus ensuring the reliability of computing power collaborative scheduling.

[0014] 4. In this invention, carbon emissions originate from grid power purchases and energy storage discharge (if high-carbon electricity is used during energy storage charging). A carbon emission traceability rule for energy storage is established: the source of electricity for energy storage charging (wind, solar, or grid) is recorded for each time period, and a charging record queue is created in chronological order; during discharge, the charging source is matched according to the principle of prioritizing green electricity, i.e., the discharged electricity is preferentially matched with previously stored green electricity, and the remaining portion is matched with grid electricity. This rule ensures the accuracy of carbon emission cost calculation and avoids deviations caused by averaging.

[0015] 5. The technical solution of this invention quantifies the extreme economic loss risk caused by the uncertainty of wind and solar forecasts through Conditional Value at Risk (CVaR), constructs a three-dimensional economic incentive function that includes electricity costs, carbon emission costs, and peak-shaving ancillary service revenue, and introduces a risk preference coefficient adjustment mechanism that is dynamically adaptive based on forecast error characteristics and market price signals. At the same time, it designs a green electricity priority traceability rule for carbon emissions from energy storage discharge. Under a unified optimization framework, it realizes the coordinated scheduling of wind and solar power, energy storage, and computing power in data centers, significantly improves the power supply reliability in extreme power shortage scenarios, reduces the overall operating cost, and for the first time incorporates peak-shaving market revenue, carbon costs, and wind and solar risks into the same decision-making model, breaking through the technical bottlenecks of traditional methods in terms of extensive risk quantification, lack of multi-market incentives, and static risk preferences.

[0016] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the method of Embodiment 1 in the present invention; Figure 2 This is a schematic diagram of the system structure in Embodiment 2 of the present invention. Detailed Implementation

[0019] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure of the invention, components and arrangements of specific examples are described below. Furthermore, reference numerals and / or letters may be repeated in different examples. This repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. It should be noted that the components illustrated in the drawings are not necessarily drawn to scale. Descriptions of well-known components, processing techniques, and processes are omitted in this invention to avoid unnecessarily limiting the invention.

[0020] Example 1 like Figure 1 As shown, this invention provides a computing-electricity collaborative scheduling method that integrates risk quantification and multiple incentives, including: Step S1: Obtain multi-source data of the region where the data center is located; the multi-source data includes wind and solar forecast data, carbon price signals, electricity price signals, peak shaving compensation prices, standby capacity prices, computing power task queues, and energy storage status. Step S2: Based on multi-source data, the conditional value at risk method is used to calculate the expected economic losses that may be caused by wind and solar forecast errors in each future period. Step S3: Adaptively adjust the risk preference coefficient based on recent forecast error characteristics, carbon price signals, and peak-shaving compensation prices; Step S4: Based on the expected economic losses that may be caused by wind and solar forecast errors in future time periods and the adjusted risk preference coefficient, construct a comprehensive cost objective function; Step S5: With the goal of minimizing the total cost, solve the comprehensive cost objective function under preset constraints to obtain and output the optimal scheduling scheme.

[0021] In step S1, data is collected at a frequency ranging from minutes to hours, from the following four types of core data: 1. Wind and solar resource data: collected through meteorological monitoring networks and power dispatching data networks. Numerical Weather Prediction (NWP) Data: Through the internal meteorological server, we access the data interface of the China Meteorological Administration to obtain high-precision gridded meteorological data for the next 72 hours or more, including elements such as cloud cover, air pressure, solar radiation, wind speed, wind direction, temperature, and humidity.

[0022] Real-time meteorological monitoring data: Through a network of meteorological sensors (including environmental monitoring instruments, wind towers, and irradiance meters) deployed in photovoltaic power plants and wind farms, data such as ambient temperature and humidity, wind speed and direction, total solar radiation (horizontal and inclined planes), direct radiation, and diffuse radiation are collected in real time.

[0023] Satellite remote sensing and cloud image data: Satellite cloud image data is acquired through satellite ground receiving stations or meteorological service provider interfaces for cloud movement prediction and large-scale weather trend analysis.

[0024] SCADA system operation data: Real-time reading of inverter and string power output, tilt angle, azimuth angle, shading status, and wind turbine speed, pitch angle, power generation, and other operating parameters from the SCADA (data acquisition and monitoring control) systems of photovoltaic power plants and wind farms.

[0025] Historical power output data: Read the actual power output records of wind and solar power over a period of time from the historical database of the SCADA system, which are used for error distribution modeling in risk quantification.

[0026] 2. Energy storage system status data: The battery management system (BMS) acquires real-time operating parameters of the energy storage system, including: SOC (State of Charge), which reflects the current remaining capacity of the energy storage; SOH (State of Health), which reflects the degree of energy storage lifespan degradation; charge / discharge power, temperature, and cycle count.

[0027] 3. Market price signals: Real-time electricity price data: Obtain spot market clearing prices from the power trading center via API interface, including medium- and long-term time-of-use prices and real-time market clearing prices.

[0028] Real-time carbon price data: Obtain the real-time transaction price of carbon allowances (CEAs) from the national carbon emission trading market or local carbon exchanges via API interface.

[0029] Peak-shaving ancillary service pricing: This includes peak-shaving compensation prices and reserve capacity prices obtained from the power dispatching authority. Peak-shaving compensation prices are typically related to response speed and response time, and are published by the electricity market.

[0030] 4. Computing power task data: Information on computing task queues is obtained from the data center's job scheduling system, including: task identifier, timeliness weight (preset or dynamically evaluated based on task type), task resource requirements (CPU / GPU core count, memory, estimated execution time), latest completion time (deadline), and task dependencies. Task queue data is collected using a subscription mechanism to obtain new tasks and status changes in real time.

[0031] The system acquires grid operation status information (peak-shaving demand signals, grid frequency) from the grid dispatch automation system and real-time electricity consumption data from metering points from the metering automation system. This data acquisition is achieved through data interfaces, API calls, or data platform services. Wind, solar, and energy storage data are acquired once per minute, market price data is acquired according to the release frequency, and computing power task data is updated in real time.

[0032] In step S2, the distribution of wind and solar forecasting errors is transformed into a quantifiable risk loss value. This represents the additional costs the system may incur under extreme conditions (such as actual wind and solar power output being far lower than forecast). The method used is Conditional Value at Risk (CVaR), the core of which is to calculate the "worst-case (1)" value. "The average loss under the β) probability scenario". Specifically, step S2 includes: S21, Define wind and solar power prediction error, which is the difference between the predicted wind and solar power output and the actual wind and solar power output. Specifically, let the wind and solar prediction error be... .in, The predicted wind and solar power output for time period t is expressed in kW. The actual wind and solar power output (unit: kW) for time period t will be collected by the SCADA system after execution. When the actual output is lower than the predicted value ( If the value is greater than 0, it requires purchasing electricity from the grid at a higher price or reducing computing power load, resulting in additional costs. Define the risk loss function L( )for:

[0033] in, The penalty cost per unit of power shortage is set at 1.2 times the peak electricity price of the grid to reflect the actual economic loss in extreme situations.

[0034] S22, Based on the wind and solar forecasting error, define a risk loss function; the risk loss function is the additional cost incurred when the actual wind and solar power output is lower than the forecasted wind and solar power output, which requires purchasing electricity from the grid at a high price or reducing the computing load. Specifically, given a confidence level β = 0.95, i.e. focusing on the worst-case scenario of 5%, the conditional value at risk (CVaR) is defined as:

[0035] in, It is the value at risk (i.e., the beta quantile of the risk loss distribution). This represents the mathematical expectation. In practical calculations, it can be achieved using historical prediction error samples or the probability distribution output by the prediction model. The risk loss function calculation of this invention includes the following key features: 1. Nonparametric estimation of error distribution The probability distribution of wind and solar forecasting errors does not pre-assume any parameterized distribution (such as a normal distribution). Instead, it directly fits the error from historical error samples using kernel density estimation (KDE). The Epanechnikov kernel function is selected, and the bandwidth follows the Silverman rule. This method can accurately capture the complex distribution characteristics of wind and solar forecasting errors, such as asymmetry, heavy tails, and multimodality, and achieves higher fitting accuracy compared to existing methods that assume the errors follow a Gaussian distribution.

[0036] 2. Latin hypercube sampling and scene reduction An initial scene set was generated using Latin Hypercube Sampling (LHS), and then the number of scenes was reduced to K representative scenes (K=1000) using Simultaneous Backward Reduction (SBR). After reduction, Wasserstein distance was used to evaluate the difference in scene distribution before and after reduction, ensuring that the reduced scene set accurately reflects the tail risk characteristics of the original distribution. This scene processing workflow ensures an optimal balance between computational accuracy and computational efficiency in CVaR calculations.

[0037] 3. Multidimensional Risk Loss Function The risk loss function of this invention It is not limited to the single cost of power outage penalties, but is constructed comprehensively from the following three dimensions: ① Direct economic loss: The extra cost of purchasing electricity from the grid at a higher price during power shortages, i.e. ,in Take 1.2 times the peak electricity price of the power grid; ② Computing Power Service Default Loss: Compensation for breach of Service Level Agreement (SLA) triggered by insufficient power supply causing interruption or delay of computing tasks. ,in As a weight for task timeliness, For task execution decision variables, The unit of computing power task breach of contract compensation coefficient (unit: yuan / task); ③ Accelerated loss of energy storage lifespan: This refers to the reduction in energy storage lifespan caused by forced deep discharge to cope with power shortages. ,in ΔSOC(t) is the weighting coefficient for energy storage lifetime loss, and ΔSOC(t) is the change in SOC.

[0038] The overall risk loss function is the sum of the above three terms, i.e. Each weighting coefficient is preset according to the data center operation strategy.

[0039] S23, Calculate the expected economic loss based on the risk loss function; specifically, the calculation of the expected economic loss is as follows:

[0040] in, Pre-set reliability for wind and solar power output prediction; Let K be the risk loss function for the k-th scenario; K is the total number of scenarios after reduction. . represents the indicator function; Let β be the quantile of the risk loss function distribution.

[0041] Specifically, for each time period t, output the risk loss value. Its meaning is: in the worst (1 β) Expected additional electricity purchase or loss costs under probabilistic scenarios (unit: yuan). The larger this value, the higher the uncertainty of the current wind and solar forecasts, and the greater the potential loss.

[0042] Preferably, the risk loss value of the present invention Instead of being calculated all at once at the start of scheduling, a rolling time-domain update mechanism is used: after each scheduling cycle (15 minutes), the latest actual wind and solar power output data is added to the historical error sample database, the error distribution is re-estimated, and the CVaR (risk loss value) for subsequent periods is updated. Simultaneously, a forgetting factor λ∈(0,1) is used to exponentially weight historical error samples, with recent error samples receiving higher weights (weight is...). ,in For the current moment, (For the sample time), the weight of the sample gradually decays in the long term, so that CVaR estimation can quickly respond to changes in weather patterns (the forgetting factor in this invention is λ=0.95).

[0043] Specifically, in step S3, the risk preference coefficient is adaptively adjusted based on recent forecast error characteristics, carbon price signals, and peak-shaving compensation prices as follows: ; ; in, This represents the risk preference coefficient for period t+1. η is the risk preference coefficient for time period t; η is the learning rate, which controls the maximum adjustment range in a single instance. A clipping function that restricts the results to a preset interval ([0,1]). t represents the change in risk preference coefficient at time t; MAPE(t) represents the average absolute percentage error (%) of the wind and solar power output prediction error over the past N time periods t. The historical average absolute percentage error of wind and solar power output prediction error (benchmark value, unit: %). The real-time carbon price for time period t (unit: yuan / kg CO2); The historical average carbon price (unit: yuan / kg CO2); The peak-shaving compensation price for time period t (unit: yuan / kWh); The historical average price for peak shaving compensation (unit: yuan / kWh); , , These are the wind and solar power output prediction error, carbon price, and peak-shaving compensation price sensitivity coefficients, respectively, satisfying... + + =1, take respectively , , ; Given a hyperbolic tangent function, map the adjustment amount to ( The interval is 1,1).

[0044] Specifically, the adjustment of the risk preference coefficient θ(t) does not employ a discrete if-then rule, but is instead uniformly calculated through the following continuous dynamic adjustment function. When A value greater than 0 indicates an increased risk level. It should be adjusted upwards (tending towards conservatism); when A value less than 0 indicates a low current risk. It should be lowered (becoming more aggressive).

[0045] Preferably, the adaptive adjustment of the risk preference coefficient based on recent forecast error characteristics, carbon price signals, and peak-shaving compensation prices further includes: We introduce higher-order statistics of the prediction error distribution as auxiliary adjustment signals to calculate the skewness S(t) and kurtosis K(t) of the prediction error over the past N time periods. If skewness If the skewness or kurtosis K(t) is greater than the first preset skewness threshold, the change in the risk preference coefficient also includes the basic adjustment increase item; When skewness If the skewness K(t) is greater than the second preset skewness threshold or the kurtosis K(t) is greater than the second preset kurtosis threshold, the change in the risk preference coefficient also includes the basic adjustment increase item and the additional adjustment increase item; wherein, the second preset skewness threshold is greater than the first preset skewness threshold, and the second preset kurtosis threshold is greater than the first preset kurtosis threshold. When the prediction error exceeds the set error threshold for N1 consecutive time periods, the risk preference coefficient is forcibly set to 1.

[0046] Specifically, skewness This reflects the asymmetry of the error distribution. A value >0 (the first preset skewness threshold) indicates a systematic underestimation of wind and solar power output (actual output is often lower than forecast), and should be adjusted upwards. ; kurtosis The tail thickness reflects the error distribution. A larger value (the second preset kurtosis threshold, for example, 2) indicates a higher probability of extreme errors occurring, and should be adjusted upwards. .

[0047] when >0.5 (second preset skewness threshold) or When the threshold is >4 (second preset kurtosis threshold) (relative to normal distribution K=3), an additional adjustment is made on top of the base adjustment. In order to address the increasing tail risks.

[0048]

[0049] in, Add items to the basic adjustment; Add items for additional adjustments.

[0050] Emergency mode trigger: When the prediction error exceeds the set threshold for 6 consecutive time periods (i.e., 1.5 hours), θ is forcibly set to 1, and the emergency lower limit of the energy storage SOC is raised to 30%, prohibiting deep discharge of the energy storage. This rule is independent of the aforementioned continuous adjustment function and serves as a safety protection mechanism in extreme situations.

[0051] In step S4, electricity costs, carbon emission costs, and peak-shaving ancillary service revenue are unified into a single comprehensive cost objective function J. The scheduling system will make decisions with the objective of minimizing J. This is the first time that peak-shaving ancillary service revenue has been used as an independent incentive dimension, deeply coupled with wind and solar risk quantification and computing power scheduling. The calculated risk loss value... Based on risk preference coefficient After weighting, add it to the objective function and implement it. Adaptive adjustment. The adjustment is based on real-time prediction error characteristics and changes in carbon price and peak-shaving price, which solves the defect of fixed risk preference in existing technologies.

[0052] The comprehensive cost objective function is as follows: ; in, For comprehensive costs; Let t be the electricity cost during the time period. The carbon emission cost for time period t; For peak-shaving ancillary service revenue during time period t; The risk preference coefficient for time period t; Let t be the expected economic loss for period t; T is the total number of periods.

[0053] Specifically, the electricity cost during time period t The specific calculation method is as follows:

[0054] in, The power purchased from the grid during time period t; The duration of time period t; The real-time electricity price for time period t; Carbon emission costs during period t The specific calculation method is as follows:

[0055]

[0056] in, The real-time carbon price (yuan / kg CO2) for time period t. Let t be the equivalent carbon emissions during the time period. Carbon emission factor of power grid (kg CO2 / kWh); (t) represents the discharge power of the non-green electricity portion in the energy storage discharge; Average carbon emission factor when charging energy storage; Specifically, carbon emissions originate from grid power purchases and energy storage discharges (if high-carbon electricity is used during energy storage charging). A carbon emission traceability rule is established for energy storage: the system records the source of electricity used for energy storage charging (wind, solar, or grid) for each time period and establishes a "charging record queue" in chronological order; during discharge, the charging source is matched according to the "green electricity priority" principle, that is, the discharged electricity is preferentially matched with previously stored green electricity, and the remaining portion is matched with grid electricity. This rule ensures the accuracy of carbon emission cost calculations and avoids deviations caused by averaging.

[0057] Peak shaving ancillary service revenue during period t The specific calculation method is as follows:

[0058] in, Indicates the peak-shaving compensation price; This represents the actual reduction in grid power purchases during time period t; This represents the price of spare capacity during time period t; This represents the reserve capacity during time period t.

[0059] Specifically, revenue from peak shaving ancillary services Data centers can participate in two types of ancillary services: Peak shaving service (peak reduction response): Actively reducing electricity load during the period specified by the power grid to obtain compensation.

[0060]

[0061] in This represents the actual reduction in grid power purchases during period t (relative to the baseline load).

[0062] Standby capacity service: Request interruptible capacity from the power grid in advance and receive a capacity fee.

[0063]

[0064] in The reserved capacity (kW) committed for time period t.

[0065] Total ancillary service revenue:

[0066] in, This represents the peak-shaving compensation price for time period t, expressed in yuan / kWh. This represents the actual reduction in grid power purchases during time period t (relative to baseline load), in kW. This indicates the price of standby capacity during time period t, in yuan / kW. This represents the committed reserve capacity for time period t, expressed in kW.

[0067] In step S5, the objective function and various physical and market constraints from step S4 are transformed into an optimization problem and solved. Mixed-integer linear programming (MILP) is employed.

[0068] 1. Constraints This includes power balance, energy storage SOC dynamics, SOC safety boundaries, task delay constraints, and grid power extraction limits, and adds the following market-related constraints: (1) Power balance constraint

[0069] , Actual output (kW) of wind and solar power is constrained by the prediction confidence interval.

[0070] , Energy storage discharge and charging power (kW), non-negative.

[0071] Total data center load (kW) is determined by the allocation of computing power tasks.

[0072] (2) Dynamic constraints of energy storage SOC

[0073] State of charge (0~1) at the end of time period t+1 : The state of charge (0~1) at the end of time period t.

[0074] Rated energy storage capacity (kWh).

[0075] , Energy storage discharge and charging efficiency (0~1).

[0076] (3) SOC security boundary

[0077] In emergency mode, the lower limit is raised to 0.3.

[0078] (4) Uncertainty constraints of wind and solar power

[0079] The lower confidence limit (kW) of the wind and solar forecast for time period t is taken as the 90th or 95th percentile to ensure conservative scheduling.

[0080] (5) Task delay constraints (for each task i)

[0081] : Decision variable, whether task i is executed during time period t.

[0082] : The latest completion time slot number (integer) for task i.

[0083] (6) Maximum power consumption from the power grid The upper limit (kW) of the power purchase agreement between the data center and the power grid.

[0084] (7) Mutual exclusion of energy storage charging and discharging Implement this by introducing binary variables y(t)∈{0,1}:

[0085] A sufficiently large constant (which can be twice the rated power of the energy storage).

[0086] Indicates charging. This indicates a discharge.

[0087] (8) Feasibility constraints of peak-shaving response:

[0088] in: : The declared peak shaving and reduction power (kW).

[0089] : The timeliness weight of task i (0~1), the higher the weight, the less likely it is to be delayed.

[0090] Task Delay Factor (dimensionless): This represents the percentage of power that can be reduced when a task is delayed. The default value is 0.8.

[0091] The additional power (kW) that energy storage can discharge relative to the baseline, i.e., the extra discharge capacity that energy storage can provide by changing the charge and discharge plan, must meet the energy storage capacity constraint.

[0092] (9) Reserve capacity constraints:

[0093] in: : The declared reserve capacity (kW).

[0094] Maximum energy storage discharge power (kW).

[0095] Minimum allowed SOC (default 0.1, emergency mode 0.3).

[0096] (10) Constraints on the carbon emission tracing queue for energy storage: The system maintains a FIFO (First-In, First-Out) queue, recording the "energy consumption - carbon emission factor" pair for each charge. During discharge, a record is retrieved from the head of the queue and matched to ensure that the carbon emission factor of the discharged energy matches the charging source. Specific rules: During charging: Charge the amount of electricity. and its carbon emission factors Add to the end of the queue. If the charging source is wind or solar power, then... If it comes from the power grid, then (Electricity emission factor of power grid).

[0097] During discharge: Records are retrieved sequentially from the head of the queue, and the carbon emission factor of each record is used to calculate the carbon emissions of the discharged portion. If the head battery is insufficient, continue recording subsequent data until the required discharge level is reached.

[0098] This queue constraint is implicit in the calculation of carbon emission costs, ensuring The energy storage and discharge portion of the device has a traceable source.

[0099] 2. Solution and Output Input the objective function and constraints into the solver in standard MILP format. The solver uses the branch and bound (B&B) method to obtain an exact solution and outputs the optimal scheduling scheme and details of revenue for each market.

[0100] The MILP problem of this invention includes the following decision variables: Integer variables: Task assignment variables (When each task is executed), charge / discharge mutual exclusion variables ; Continuous variable: Energy storage charging and discharging power Power purchased by the power grid Peak shaving and power reduction standby capacity .

[0101] Branch and bound is the core algorithm framework of all current mainstream MILP solvers (Gurobi, CPLEX, SCIP). Its basic idea is to recursively divide the feasible solution space of the original problem into increasingly smaller subsets (branches), and calculate the lower bound (bound) of the objective function value for each subset, thereby pruning subsets that cannot contain the optimal solution and narrowing the search range. The algorithm execution flow is as follows: Step 1: Presolving Before the formal solution is obtained, the solver preprocesses the original problem: Redundant constraint elimination: Remove constraints that are implicitly included by other constraints; Variable boundary tightening: Narrowing the upper and lower bounds of variables based on constraints; Fixed variable: If a variable takes the same value in all feasible solutions, then it is directly fixed; Testing infeasible: If the constraints are contradictory, terminate the test early.

[0102] Pre-solving can significantly reduce the problem dimensionality and shorten the solution time.

[0103] Step 2: Solve the root node LP relaxation problem Temporarily relax all integer variable constraints in the original MILP problem to continuous variables (e.g., ... ∈{0,1} relaxes to 0≤ ≤1), resulting in a linear programming (LP) relaxation problem. The solver uses the dual simplex method to solve this relaxation problem: If the LP relaxation problem has no feasible solution, then the original MILP problem also has no feasible solution, and the algorithm terminates. If all integer variables in the optimal solution of the LP relaxation problem are integers, then the solution is the optimal solution of the original MILP problem, and the algorithm terminates. If the optimal solution to the LP relaxation problem contains integer variables with non-integer values ​​(such as a certain...), If the value is 0.7, then proceed to the branch step.

[0104] The optimal objective function value of the root node LP constitutes the lower bound (LB) of the original problem—that is, the objective function value of the optimal integer solution cannot be lower than this value (for minimization problems).

[0105] Step 3: Branching Starting from the root node, select an integer variable whose value is not an integer to branch from. Specifically, if the variable... The relaxed solution is x =0.7 (not an integer), then two subproblems are created: Left branch: Add constraints ≤0 (forces the task not to be executed during this time period); Right branch: Add constraints ≥1 (Force this task to be executed during this period).

[0106] The feasible regions of these two subproblems are disjoint and cover all integer feasible solutions to the original problem. The branching variable selection strategies include strong branching (trying all candidate variables in advance and selecting the best one), pseudo-cost branching (using historical information to estimate the improvement of the lower bound after branching each variable), and least feasible branching (selecting the variable whose fractional part is closest to 0.5).

[0107] Step 4: Bounding For each newly generated subproblem, solve its LP relaxation problem again to obtain the lower bound of the subproblem. At the same time, if during the branching process it is found that the LP relaxation solution of a subproblem exactly satisfies all integer constraints (i.e. all integer variables are integers), then update the global upper bound (UB), which is the currently found optimal integer feasible solution.

[0108] Step 5: Pruning Based on the bounding results, the subproblems are pruned: Infeasible pruning: If the LP relaxation of a subproblem has no feasible solution, the subproblem cannot contain feasible integer solutions and should be discarded directly. Integer solution pruning: If the LP relaxation solution of a subproblem is already an integer feasible solution, then the subproblem has found the optimal solution and there is no need to continue branching; Lower bound pruning: If the LP relaxation lower bound of a subproblem is not lower than the current global upper bound (LB≥UB), then the subproblem cannot contain a better solution and should be discarded directly.

[0109] Pruning is key to the efficiency of branch and bound algorithms—it avoids exhaustively searching for all integer combinations.

[0110] Step 6: Node Selection and Iteration Choose one from all active nodes (subproblems that have not yet been pruned or explored) to proceed to the next step. Node selection strategies include depth-first (prioritizing the most recently generated subproblems to quickly obtain the upper bound) and best-first (prioritizing the subproblems with the smallest lower bound, which are theoretically most likely to contain the optimal solution).

[0111] Iterate through steps 3 to 6 until the list of active nodes is empty or other termination conditions are met.

[0112] Step 7: Output the optimal solution When the algorithm terminates: If at least one integer feasible solution is found, then the solution corresponding to the global upper bound is the optimal solution; If no integer feasible solution is found, then the original MILP problem is not feasible on integers.

[0113] The solver outputs the values ​​of all decision variables in the optimal solution, including: energy storage charging and discharging power at each time period. The execution time of each task ( =1 corresponds to t), and the power purchased by the power grid. The declared peak shaving and reduction power and spare capacity and the corresponding minimum total cost Detailed breakdown of costs and benefits for each item.

[0114] Combine, return The minimum solution. Output includes: Energy storage charging and discharging power at each time period , .

[0115] The execution time of each task ( The corresponding t).

[0116] Power purchased by the power grid .

[0117] Reported peak shaving and reduction power and spare capacity .

[0118] Detailed breakdown of revenue for each market (ancillary service revenue, carbon emission costs, electricity costs).

[0119] 4. Execution in the rolling time domain Since the MILP solution time increases with the problem size, a rolling time-domain strategy is adopted: each time, only the next T=4 time periods (1 hour) are solved. The scheduling instructions for the first time period are executed, and then the time window is shifted forward by one time period to resolve using the latest wind, solar, energy storage, and task data. This process is repeated to achieve real-time optimization control.

[0120] Preferably, the computing-electricity collaborative scheduling method integrating risk quantification and multiple incentives provided by the technical solution of the present invention further includes: Step S6: Execute the optimal scheduling scheme, update the error prediction characteristics based on the actual execution results of the optimal scheduling scheme, and adjust the risk preference coefficient based on the updated prediction error characteristics, carbon price signal, and peak shaving compensation price.

[0121] Specifically, after issuing energy storage charging and discharging instructions and computing power task scheduling instructions, the declared peak shaving / reserve capacity is submitted to the power market platform. The actual wind and solar power output, grid power purchase, and energy storage SOC changes are collected. The deviation between the actual response and the declared amount is calculated. If the deviation exceeds the threshold, the penalty cost is recorded (for the next scheduling correction), and the prediction error characteristics are updated for the next round of θ adjustment.

[0122] Preferably, steps S2 to S6 are repeated, and each scheduling cycle can be executed in a rolling manner to achieve real-time adaptive optimization.

[0123] The technical solution of this invention acquires multi-source data of the region where the data center is located. This multi-source data includes wind and solar forecast data, carbon price signals, electricity price signals, peak-shaving compensation prices, reserve capacity prices, computing power task queues, and energy storage status. Based on this multi-source data, the conditional value at risk method is used to calculate the expected economic losses that may result from wind and solar forecast errors in future time periods. The risk preference coefficient is adaptively adjusted based on recent forecast error characteristics, carbon price signals, and peak-shaving compensation prices. A comprehensive cost objective function is constructed based on the expected economic losses that may result from wind and solar forecast errors in future time periods and the adjusted risk preference coefficient. With the goal of minimizing the total cost, the comprehensive cost objective function is solved under preset constraints to obtain and output the optimal scheduling scheme. This effectively solves the problem of low reliability in computing-power collaborative scheduling caused by existing technologies and effectively improves the reliability of computing-power collaborative scheduling.

[0124] The technical solution of this invention introduces a risk preference coefficient, which automatically adjusts the conservative / aggressive level of risk preference based on the statistical characteristics and distribution characteristics of recent prediction errors and market price signals, thereby achieving a balance between risk and return.

[0125] In the technical solution of this invention, the risk loss function is not limited to a single power outage penalty cost, but is constructed by comprehensively considering different dimensions such as direct economic losses, computing power service default losses, and accelerated energy storage lifespan loss, thus ensuring the reliability of computing power collaborative scheduling.

[0126] In this invention, carbon emissions originate from grid power purchases and energy storage discharge (if high-carbon electricity is used during energy storage charging). A carbon emission traceability rule for energy storage is established: the source of electricity for energy storage charging (wind, solar, or grid) is recorded for each time period, and a charging record queue is created in chronological order; during discharge, the charging source is matched according to the principle of prioritizing green electricity, that is, the discharged electricity is preferentially matched with previously stored green electricity, and the remaining portion is matched with grid electricity. This rule ensures the accuracy of carbon emission cost calculation and avoids deviations caused by averaging.

[0127] This invention quantifies the extreme economic loss risk caused by the uncertainty of wind and solar forecasts using Conditional Value at Risk (CVaR). It constructs a three-dimensional economic incentive function that includes electricity costs, carbon emission costs, and peak-shaving ancillary service revenue. It also introduces a risk preference coefficient adjustment mechanism that is dynamically adaptive based on forecast error characteristics and market price signals. Simultaneously, it designs a green electricity priority traceability rule for carbon emissions from energy storage discharge. Under a unified optimization framework, it realizes the coordinated scheduling of wind and solar power, energy storage, and computing power in data centers, significantly improving the power supply reliability in extreme power shortage scenarios, reducing overall operating costs, and for the first time incorporating peak-shaving market revenue, carbon costs, and wind and solar risks into the same decision-making model. This breaks through the technical bottlenecks of traditional methods in terms of extensive risk quantification, lack of multi-market incentives, and static risk preferences.

[0128] Example 2 like Figure 2 As shown, the present invention also provides a computing-electricity collaborative scheduling system that integrates risk quantification and multiple incentives, comprising: The acquisition module acquires multi-source data from the region where the data center is located; the multi-source data includes wind and solar forecast data, carbon price signals, electricity price signals, peak-shaving compensation prices, standby capacity prices, computing power task queues, and energy storage status; The calculation module, based on multi-source data, uses the conditional value at risk method to calculate the expected economic losses that may be caused by wind and solar forecast errors in different future periods. The adjustment module adaptively adjusts the risk preference coefficient based on recent forecast error characteristics, carbon price signals, and peak-shaving compensation prices. The module constructs a comprehensive cost objective function based on the expected economic losses that may be caused by wind and solar forecast errors in future time periods and the adjusted risk preference coefficient. The solution module aims to minimize the total cost by solving the comprehensive cost objective function under preset constraints, and then obtains and outputs the optimal scheduling scheme.

[0129] The implementation methods of each module correspond to the steps in Implementation Example 1, and will not be repeated here.

[0130] The technical solution of this invention acquires multi-source data of the region where the data center is located. This multi-source data includes wind and solar forecast data, carbon price signals, electricity price signals, peak-shaving compensation prices, reserve capacity prices, computing power task queues, and energy storage status. Based on this multi-source data, the conditional value at risk method is used to calculate the expected economic losses that may result from wind and solar forecast errors in future time periods. The risk preference coefficient is adaptively adjusted based on recent forecast error characteristics, carbon price signals, and peak-shaving compensation prices. A comprehensive cost objective function is constructed based on the expected economic losses that may result from wind and solar forecast errors in future time periods and the adjusted risk preference coefficient. With the goal of minimizing the total cost, the comprehensive cost objective function is solved under preset constraints to obtain and output the optimal scheduling scheme. This effectively solves the problem of low reliability in computing-power collaborative scheduling caused by existing technologies and effectively improves the reliability of computing-power collaborative scheduling.

[0131] The technical solution of this invention introduces a risk preference coefficient, which automatically adjusts the conservative / aggressive level of risk preference based on the statistical characteristics and distribution characteristics of recent prediction errors and market price signals, thereby achieving a balance between risk and return.

[0132] In the technical solution of this invention, the risk loss function is not limited to a single power outage penalty cost, but is constructed by comprehensively considering different dimensions such as direct economic losses, computing power service default losses, and accelerated energy storage lifespan loss, thus ensuring the reliability of computing power collaborative scheduling.

[0133] In this invention, carbon emissions originate from grid power purchases and energy storage discharge (if high-carbon electricity is used during energy storage charging). A carbon emission traceability rule for energy storage is established: the source of electricity for energy storage charging (wind, solar, or grid) is recorded for each time period, and a charging record queue is created in chronological order; during discharge, the charging source is matched according to the principle of prioritizing green electricity, that is, the discharged electricity is preferentially matched with previously stored green electricity, and the remaining portion is matched with grid electricity. This rule ensures the accuracy of carbon emission cost calculation and avoids deviations caused by averaging.

[0134] This invention quantifies the extreme economic loss risk caused by the uncertainty of wind and solar forecasts using Conditional Value at Risk (CVaR). It constructs a three-dimensional economic incentive function that includes electricity costs, carbon emission costs, and peak-shaving ancillary service revenue. It also introduces a risk preference coefficient adjustment mechanism that is dynamically adaptive based on forecast error characteristics and market price signals. Simultaneously, it designs a green electricity priority traceability rule for carbon emissions from energy storage discharge. Under a unified optimization framework, it realizes the coordinated scheduling of wind and solar power, energy storage, and computing power in data centers, significantly improving the power supply reliability in extreme power shortage scenarios, reducing overall operating costs, and for the first time incorporating peak-shaving market revenue, carbon costs, and wind and solar risks into the same decision-making model. This breaks through the technical bottlenecks of traditional methods in terms of extensive risk quantification, lack of multi-market incentives, and static risk preferences.

[0135] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A computational-powered collaborative scheduling method integrating risk quantification and multiple incentives, characterized in that, include: Acquire multi-source data of the region where the data center is located; the multi-source data includes wind and solar forecast data, carbon price signals, electricity price signals, peak-shaving compensation prices, standby capacity prices, computing power task queues, and energy storage status; Based on multi-source data, the conditional value at risk method is used to calculate the expected economic losses that may be caused by wind and solar forecasting errors in different future periods. Based on recent forecast error characteristics, carbon price signals, and peak-shaving compensation prices, the risk preference coefficient is adaptively adjusted. Based on the expected economic losses that may be caused by wind and solar forecast errors in different future periods and the adjusted risk appetite coefficient, a comprehensive cost objective function is constructed. With the goal of minimizing total cost, the comprehensive cost objective function is solved under preset constraints to obtain and output the optimal scheduling scheme.

2. The computational-electricity collaborative scheduling method integrating risk quantification and multiple incentives according to claim 1, characterized in that, Based on multi-source data, the conditional value at risk approach is used to calculate the expected economic losses that may result from wind and solar forecasting errors in different future time periods, including: The wind and solar power prediction error is defined as the difference between the predicted wind and solar power output and the actual wind and solar power output. Based on the wind and solar forecasting error, a risk loss function is defined; the risk loss function is the additional cost incurred when the actual wind and solar power output is lower than the forecast value, requiring the purchase of electricity from the grid at a higher price or the reduction of computing power load. Calculate the expected economic loss based on the risk loss function; specifically, the calculation of the expected economic loss is as follows: in, Pre-set reliability for wind and solar power output prediction; Let K be the risk loss function for the k-th scenario; K is the total number of scenarios after reduction. . represents the indicator function; Let β be the quantile of the risk loss function distribution.

3. The computational-electricity collaborative scheduling method integrating risk quantification and multiple incentives according to claim 2, characterized in that, The risk loss function is calculated as follows: ， in, Let t be the risk loss function for time period t; The wind and solar forecast error for time period t; The weighting coefficient for direct economic losses caused by wind and solar forecasting errors; The penalty cost per unit of power outage during time period t; Let i be the timeliness weight for task i; Let i be the execution decision variable for task i. The default compensation coefficient per unit of computing power task; ΔSOC(t) represents the weighting coefficient for the lifespan loss of the energy storage system, and ΔSOC(t) represents the change in SOC of the energy storage system during time period t.

4. The computational-electricity collaborative scheduling method integrating risk quantification and multiple incentives according to claim 1, characterized in that, Based on recent forecast error characteristics, carbon price signals, and peak-shaving compensation prices, the risk appetite coefficient is adaptively adjusted as follows: ； ； in, This represents the risk preference coefficient for period t+1. η is the risk preference coefficient for time period t; η is the learning rate. A clipping function that restricts the results to a preset range; t represents the change in risk preference coefficient at time t; MAPE(t) represents the average absolute percentage error of the wind and solar power output prediction error over the past N time periods t. The historical average absolute percentage error of wind and solar power output prediction error; The real-time carbon price for time period t; This represents the historical average carbon price. The peak-shaving compensation price for time period t; The peak-shaving compensation price is the historical average. , , These are the wind and solar power output prediction error, carbon price, and peak-shaving compensation price sensitivity coefficient, respectively. It is the hyperbolic tangent function.

5. The computational-electricity collaborative scheduling method integrating risk quantification and multiple incentives according to claim 4, characterized in that, Based on recent forecast error characteristics, carbon price signals, and peak-shaving compensation prices, the adaptive adjustment of the risk appetite coefficient also includes: We introduce higher-order statistics of the prediction error distribution as auxiliary adjustment signals to calculate the skewness S(t) and kurtosis K(t) of the prediction error over the past N time periods. If skewness If the skewness or kurtosis K(t) is greater than the first preset skewness threshold, the change in the risk preference coefficient also includes the basic adjustment increase item; When skewness If the skewness K(t) is greater than the second preset skewness threshold or the kurtosis K(t) is greater than the second preset kurtosis threshold, the change in the risk preference coefficient also includes a basic adjustment increase item and an additional adjustment increase item; wherein, the second preset skewness threshold is greater than the first preset skewness threshold, and the second preset kurtosis threshold is greater than the first preset kurtosis threshold. When the prediction error exceeds the set error threshold for N1 consecutive time periods, the risk preference coefficient is forcibly set to 1.

6. The computational-electricity collaborative scheduling method integrating risk quantification and multiple incentives according to claim 1, characterized in that, The comprehensive cost objective function is as follows: ； in, For comprehensive costs; Let t be the electricity cost during the time period. The carbon emission cost for time period t; For peak-shaving ancillary service revenue during time period t; The risk preference coefficient for time period t; Let t be the expected economic loss for period t; T is the total number of periods.

7. The computational-electricity collaborative scheduling method integrating risk quantification and multiple incentives according to claim 6, characterized in that, Electricity cost during period t The specific calculation method is as follows: in, The power purchased from the grid during time period t; The duration of time period t; The real-time electricity price for time period t; Carbon emission costs during period t The specific calculation method is as follows: in, The real-time carbon price for time period t; Let t be the equivalent carbon emissions during the time period. Carbon emission factor of power grid; (t) represents the discharge power of the non-green electricity portion in the energy storage discharge; Average carbon emission factor when charging energy storage; Peak shaving ancillary service revenue during period t The specific calculation method is as follows: in, Indicates the peak-shaving compensation price; This represents the actual reduction in grid power purchases during time period t; This represents the price of spare capacity during time period t; This represents the reserve capacity during time period t.

8. The computational-electricity collaborative scheduling method integrating risk quantification and multiple incentives according to claim 1, characterized in that, The preset constraints include power balance constraints, energy storage SOC dynamic constraints, SOC safety boundary constraints, task delay constraints, grid power draw upper limit constraints, energy storage charging and discharging mutual exclusion constraints, peak shaving response feasibility constraints, reserve capacity reservation constraints, and energy storage carbon emission tracing queue constraints.

9. A computational-electricity collaborative scheduling method integrating risk quantification and multiple incentives according to any one of claims 1-8, characterized in that, Also includes: Execute the optimal scheduling scheme, update the error prediction characteristics based on the actual execution results of the optimal scheduling scheme, and adjust the risk preference coefficient based on the updated prediction error characteristics, carbon price signal, and peak shaving compensation price.

10. A computer-based collaborative scheduling system integrating risk quantification and multiple incentives, characterized in that, include: The acquisition module acquires multi-source data from the region where the data center is located; the multi-source data includes wind and solar forecast data, carbon price signals, electricity price signals, peak-shaving compensation prices, standby capacity prices, computing power task queues, and energy storage status; The calculation module, based on multi-source data, uses the conditional value at risk method to calculate the expected economic losses that may be caused by wind and solar forecast errors in different future periods. The adjustment module adaptively adjusts the risk preference coefficient based on recent forecast error characteristics, carbon price signals, and peak-shaving compensation prices. The module constructs a comprehensive cost objective function based on the expected economic losses that may be caused by wind and solar forecast errors in future time periods and the adjusted risk preference coefficient. The solution module aims to minimize the total cost by solving the comprehensive cost objective function under preset constraints, and then obtains and outputs the optimal scheduling scheme.

Images